1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This transformation analyzes and transforms the induction variables (and 11 // computations derived from them) into simpler forms suitable for subsequent 12 // analysis and transformation. 13 // 14 // If the trip count of a loop is computable, this pass also makes the following 15 // changes: 16 // 1. The exit condition for the loop is canonicalized to compare the 17 // induction value against the exit value. This turns loops like: 18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)' 19 // 2. Any use outside of the loop of an expression derived from the indvar 20 // is changed to compute the derived value outside of the loop, eliminating 21 // the dependence on the exit value of the induction variable. If the only 22 // purpose of the loop is to compute the exit value of some derived 23 // expression, this transformation will make the loop dead. 24 // 25 //===----------------------------------------------------------------------===// 26 27 #define DEBUG_TYPE "indvars" 28 #include "llvm/Transforms/Scalar.h" 29 #include "llvm/ADT/DenseMap.h" 30 #include "llvm/ADT/SmallVector.h" 31 #include "llvm/ADT/Statistic.h" 32 #include "llvm/Analysis/Dominators.h" 33 #include "llvm/Analysis/LoopInfo.h" 34 #include "llvm/Analysis/LoopPass.h" 35 #include "llvm/Analysis/ScalarEvolutionExpander.h" 36 #include "llvm/IR/BasicBlock.h" 37 #include "llvm/IR/Constants.h" 38 #include "llvm/IR/DataLayout.h" 39 #include "llvm/IR/Instructions.h" 40 #include "llvm/IR/IntrinsicInst.h" 41 #include "llvm/IR/LLVMContext.h" 42 #include "llvm/IR/Type.h" 43 #include "llvm/Support/CFG.h" 44 #include "llvm/Support/CommandLine.h" 45 #include "llvm/Support/Debug.h" 46 #include "llvm/Support/raw_ostream.h" 47 #include "llvm/Target/TargetLibraryInfo.h" 48 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 49 #include "llvm/Transforms/Utils/Local.h" 50 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 51 using namespace llvm; 52 53 STATISTIC(NumWidened , "Number of indvars widened"); 54 STATISTIC(NumReplaced , "Number of exit values replaced"); 55 STATISTIC(NumLFTR , "Number of loop exit tests replaced"); 56 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); 57 STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); 58 59 // Trip count verification can be enabled by default under NDEBUG if we 60 // implement a strong expression equivalence checker in SCEV. Until then, we 61 // use the verify-indvars flag, which may assert in some cases. 62 static cl::opt<bool> VerifyIndvars( 63 "verify-indvars", cl::Hidden, 64 cl::desc("Verify the ScalarEvolution result after running indvars")); 65 66 namespace { 67 class IndVarSimplify : public LoopPass { 68 LoopInfo *LI; 69 ScalarEvolution *SE; 70 DominatorTree *DT; 71 DataLayout *TD; 72 TargetLibraryInfo *TLI; 73 74 SmallVector<WeakVH, 16> DeadInsts; 75 bool Changed; 76 public: 77 78 static char ID; // Pass identification, replacement for typeid 79 IndVarSimplify() : LoopPass(ID), LI(0), SE(0), DT(0), TD(0), 80 Changed(false) { 81 initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry()); 82 } 83 84 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 85 86 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 87 AU.addRequired<DominatorTree>(); 88 AU.addRequired<LoopInfo>(); 89 AU.addRequired<ScalarEvolution>(); 90 AU.addRequiredID(LoopSimplifyID); 91 AU.addRequiredID(LCSSAID); 92 AU.addPreserved<ScalarEvolution>(); 93 AU.addPreservedID(LoopSimplifyID); 94 AU.addPreservedID(LCSSAID); 95 AU.setPreservesCFG(); 96 } 97 98 private: 99 virtual void releaseMemory() { 100 DeadInsts.clear(); 101 } 102 103 bool isValidRewrite(Value *FromVal, Value *ToVal); 104 105 void HandleFloatingPointIV(Loop *L, PHINode *PH); 106 void RewriteNonIntegerIVs(Loop *L); 107 108 void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM); 109 110 void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); 111 112 Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount, 113 PHINode *IndVar, SCEVExpander &Rewriter); 114 115 void SinkUnusedInvariants(Loop *L); 116 }; 117 } 118 119 char IndVarSimplify::ID = 0; 120 INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars", 121 "Induction Variable Simplification", false, false) 122 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 123 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 124 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 125 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 126 INITIALIZE_PASS_DEPENDENCY(LCSSA) 127 INITIALIZE_PASS_END(IndVarSimplify, "indvars", 128 "Induction Variable Simplification", false, false) 129 130 Pass *llvm::createIndVarSimplifyPass() { 131 return new IndVarSimplify(); 132 } 133 134 /// isValidRewrite - Return true if the SCEV expansion generated by the 135 /// rewriter can replace the original value. SCEV guarantees that it 136 /// produces the same value, but the way it is produced may be illegal IR. 137 /// Ideally, this function will only be called for verification. 138 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) { 139 // If an SCEV expression subsumed multiple pointers, its expansion could 140 // reassociate the GEP changing the base pointer. This is illegal because the 141 // final address produced by a GEP chain must be inbounds relative to its 142 // underlying object. Otherwise basic alias analysis, among other things, 143 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid 144 // producing an expression involving multiple pointers. Until then, we must 145 // bail out here. 146 // 147 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject 148 // because it understands lcssa phis while SCEV does not. 149 Value *FromPtr = FromVal; 150 Value *ToPtr = ToVal; 151 if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) { 152 FromPtr = GEP->getPointerOperand(); 153 } 154 if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) { 155 ToPtr = GEP->getPointerOperand(); 156 } 157 if (FromPtr != FromVal || ToPtr != ToVal) { 158 // Quickly check the common case 159 if (FromPtr == ToPtr) 160 return true; 161 162 // SCEV may have rewritten an expression that produces the GEP's pointer 163 // operand. That's ok as long as the pointer operand has the same base 164 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the 165 // base of a recurrence. This handles the case in which SCEV expansion 166 // converts a pointer type recurrence into a nonrecurrent pointer base 167 // indexed by an integer recurrence. 168 169 // If the GEP base pointer is a vector of pointers, abort. 170 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy()) 171 return false; 172 173 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr)); 174 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr)); 175 if (FromBase == ToBase) 176 return true; 177 178 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out " 179 << *FromBase << " != " << *ToBase << "\n"); 180 181 return false; 182 } 183 return true; 184 } 185 186 /// Determine the insertion point for this user. By default, insert immediately 187 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the 188 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest 189 /// common dominator for the incoming blocks. 190 static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 191 DominatorTree *DT) { 192 PHINode *PHI = dyn_cast<PHINode>(User); 193 if (!PHI) 194 return User; 195 196 Instruction *InsertPt = 0; 197 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 198 if (PHI->getIncomingValue(i) != Def) 199 continue; 200 201 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 202 if (!InsertPt) { 203 InsertPt = InsertBB->getTerminator(); 204 continue; 205 } 206 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 207 InsertPt = InsertBB->getTerminator(); 208 } 209 assert(InsertPt && "Missing phi operand"); 210 assert((!isa<Instruction>(Def) || 211 DT->dominates(cast<Instruction>(Def), InsertPt)) && 212 "def does not dominate all uses"); 213 return InsertPt; 214 } 215 216 //===----------------------------------------------------------------------===// 217 // RewriteNonIntegerIVs and helpers. Prefer integer IVs. 218 //===----------------------------------------------------------------------===// 219 220 /// ConvertToSInt - Convert APF to an integer, if possible. 221 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { 222 bool isExact = false; 223 // See if we can convert this to an int64_t 224 uint64_t UIntVal; 225 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, 226 &isExact) != APFloat::opOK || !isExact) 227 return false; 228 IntVal = UIntVal; 229 return true; 230 } 231 232 /// HandleFloatingPointIV - If the loop has floating induction variable 233 /// then insert corresponding integer induction variable if possible. 234 /// For example, 235 /// for(double i = 0; i < 10000; ++i) 236 /// bar(i) 237 /// is converted into 238 /// for(int i = 0; i < 10000; ++i) 239 /// bar((double)i); 240 /// 241 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) { 242 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 243 unsigned BackEdge = IncomingEdge^1; 244 245 // Check incoming value. 246 ConstantFP *InitValueVal = 247 dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); 248 249 int64_t InitValue; 250 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) 251 return; 252 253 // Check IV increment. Reject this PN if increment operation is not 254 // an add or increment value can not be represented by an integer. 255 BinaryOperator *Incr = 256 dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); 257 if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return; 258 259 // If this is not an add of the PHI with a constantfp, or if the constant fp 260 // is not an integer, bail out. 261 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1)); 262 int64_t IncValue; 263 if (IncValueVal == 0 || Incr->getOperand(0) != PN || 264 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue)) 265 return; 266 267 // Check Incr uses. One user is PN and the other user is an exit condition 268 // used by the conditional terminator. 269 Value::use_iterator IncrUse = Incr->use_begin(); 270 Instruction *U1 = cast<Instruction>(*IncrUse++); 271 if (IncrUse == Incr->use_end()) return; 272 Instruction *U2 = cast<Instruction>(*IncrUse++); 273 if (IncrUse != Incr->use_end()) return; 274 275 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't 276 // only used by a branch, we can't transform it. 277 FCmpInst *Compare = dyn_cast<FCmpInst>(U1); 278 if (!Compare) 279 Compare = dyn_cast<FCmpInst>(U2); 280 if (Compare == 0 || !Compare->hasOneUse() || 281 !isa<BranchInst>(Compare->use_back())) 282 return; 283 284 BranchInst *TheBr = cast<BranchInst>(Compare->use_back()); 285 286 // We need to verify that the branch actually controls the iteration count 287 // of the loop. If not, the new IV can overflow and no one will notice. 288 // The branch block must be in the loop and one of the successors must be out 289 // of the loop. 290 assert(TheBr->isConditional() && "Can't use fcmp if not conditional"); 291 if (!L->contains(TheBr->getParent()) || 292 (L->contains(TheBr->getSuccessor(0)) && 293 L->contains(TheBr->getSuccessor(1)))) 294 return; 295 296 297 // If it isn't a comparison with an integer-as-fp (the exit value), we can't 298 // transform it. 299 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1)); 300 int64_t ExitValue; 301 if (ExitValueVal == 0 || 302 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue)) 303 return; 304 305 // Find new predicate for integer comparison. 306 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE; 307 switch (Compare->getPredicate()) { 308 default: return; // Unknown comparison. 309 case CmpInst::FCMP_OEQ: 310 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break; 311 case CmpInst::FCMP_ONE: 312 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break; 313 case CmpInst::FCMP_OGT: 314 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break; 315 case CmpInst::FCMP_OGE: 316 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break; 317 case CmpInst::FCMP_OLT: 318 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break; 319 case CmpInst::FCMP_OLE: 320 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break; 321 } 322 323 // We convert the floating point induction variable to a signed i32 value if 324 // we can. This is only safe if the comparison will not overflow in a way 325 // that won't be trapped by the integer equivalent operations. Check for this 326 // now. 327 // TODO: We could use i64 if it is native and the range requires it. 328 329 // The start/stride/exit values must all fit in signed i32. 330 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue)) 331 return; 332 333 // If not actually striding (add x, 0.0), avoid touching the code. 334 if (IncValue == 0) 335 return; 336 337 // Positive and negative strides have different safety conditions. 338 if (IncValue > 0) { 339 // If we have a positive stride, we require the init to be less than the 340 // exit value. 341 if (InitValue >= ExitValue) 342 return; 343 344 uint32_t Range = uint32_t(ExitValue-InitValue); 345 // Check for infinite loop, either: 346 // while (i <= Exit) or until (i > Exit) 347 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) { 348 if (++Range == 0) return; // Range overflows. 349 } 350 351 unsigned Leftover = Range % uint32_t(IncValue); 352 353 // If this is an equality comparison, we require that the strided value 354 // exactly land on the exit value, otherwise the IV condition will wrap 355 // around and do things the fp IV wouldn't. 356 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 357 Leftover != 0) 358 return; 359 360 // If the stride would wrap around the i32 before exiting, we can't 361 // transform the IV. 362 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue) 363 return; 364 365 } else { 366 // If we have a negative stride, we require the init to be greater than the 367 // exit value. 368 if (InitValue <= ExitValue) 369 return; 370 371 uint32_t Range = uint32_t(InitValue-ExitValue); 372 // Check for infinite loop, either: 373 // while (i >= Exit) or until (i < Exit) 374 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) { 375 if (++Range == 0) return; // Range overflows. 376 } 377 378 unsigned Leftover = Range % uint32_t(-IncValue); 379 380 // If this is an equality comparison, we require that the strided value 381 // exactly land on the exit value, otherwise the IV condition will wrap 382 // around and do things the fp IV wouldn't. 383 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 384 Leftover != 0) 385 return; 386 387 // If the stride would wrap around the i32 before exiting, we can't 388 // transform the IV. 389 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue) 390 return; 391 } 392 393 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext()); 394 395 // Insert new integer induction variable. 396 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN); 397 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue), 398 PN->getIncomingBlock(IncomingEdge)); 399 400 Value *NewAdd = 401 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue), 402 Incr->getName()+".int", Incr); 403 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge)); 404 405 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd, 406 ConstantInt::get(Int32Ty, ExitValue), 407 Compare->getName()); 408 409 // In the following deletions, PN may become dead and may be deleted. 410 // Use a WeakVH to observe whether this happens. 411 WeakVH WeakPH = PN; 412 413 // Delete the old floating point exit comparison. The branch starts using the 414 // new comparison. 415 NewCompare->takeName(Compare); 416 Compare->replaceAllUsesWith(NewCompare); 417 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI); 418 419 // Delete the old floating point increment. 420 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); 421 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI); 422 423 // If the FP induction variable still has uses, this is because something else 424 // in the loop uses its value. In order to canonicalize the induction 425 // variable, we chose to eliminate the IV and rewrite it in terms of an 426 // int->fp cast. 427 // 428 // We give preference to sitofp over uitofp because it is faster on most 429 // platforms. 430 if (WeakPH) { 431 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", 432 PN->getParent()->getFirstInsertionPt()); 433 PN->replaceAllUsesWith(Conv); 434 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI); 435 } 436 Changed = true; 437 } 438 439 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) { 440 // First step. Check to see if there are any floating-point recurrences. 441 // If there are, change them into integer recurrences, permitting analysis by 442 // the SCEV routines. 443 // 444 BasicBlock *Header = L->getHeader(); 445 446 SmallVector<WeakVH, 8> PHIs; 447 for (BasicBlock::iterator I = Header->begin(); 448 PHINode *PN = dyn_cast<PHINode>(I); ++I) 449 PHIs.push_back(PN); 450 451 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) 452 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i])) 453 HandleFloatingPointIV(L, PN); 454 455 // If the loop previously had floating-point IV, ScalarEvolution 456 // may not have been able to compute a trip count. Now that we've done some 457 // re-writing, the trip count may be computable. 458 if (Changed) 459 SE->forgetLoop(L); 460 } 461 462 //===----------------------------------------------------------------------===// 463 // RewriteLoopExitValues - Optimize IV users outside the loop. 464 // As a side effect, reduces the amount of IV processing within the loop. 465 //===----------------------------------------------------------------------===// 466 467 /// RewriteLoopExitValues - Check to see if this loop has a computable 468 /// loop-invariant execution count. If so, this means that we can compute the 469 /// final value of any expressions that are recurrent in the loop, and 470 /// substitute the exit values from the loop into any instructions outside of 471 /// the loop that use the final values of the current expressions. 472 /// 473 /// This is mostly redundant with the regular IndVarSimplify activities that 474 /// happen later, except that it's more powerful in some cases, because it's 475 /// able to brute-force evaluate arbitrary instructions as long as they have 476 /// constant operands at the beginning of the loop. 477 void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { 478 // Verify the input to the pass in already in LCSSA form. 479 assert(L->isLCSSAForm(*DT)); 480 481 SmallVector<BasicBlock*, 8> ExitBlocks; 482 L->getUniqueExitBlocks(ExitBlocks); 483 484 // Find all values that are computed inside the loop, but used outside of it. 485 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 486 // the exit blocks of the loop to find them. 487 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 488 BasicBlock *ExitBB = ExitBlocks[i]; 489 490 // If there are no PHI nodes in this exit block, then no values defined 491 // inside the loop are used on this path, skip it. 492 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 493 if (!PN) continue; 494 495 unsigned NumPreds = PN->getNumIncomingValues(); 496 497 // Iterate over all of the PHI nodes. 498 BasicBlock::iterator BBI = ExitBB->begin(); 499 while ((PN = dyn_cast<PHINode>(BBI++))) { 500 if (PN->use_empty()) 501 continue; // dead use, don't replace it 502 503 // SCEV only supports integer expressions for now. 504 if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy()) 505 continue; 506 507 // It's necessary to tell ScalarEvolution about this explicitly so that 508 // it can walk the def-use list and forget all SCEVs, as it may not be 509 // watching the PHI itself. Once the new exit value is in place, there 510 // may not be a def-use connection between the loop and every instruction 511 // which got a SCEVAddRecExpr for that loop. 512 SE->forgetValue(PN); 513 514 // Iterate over all of the values in all the PHI nodes. 515 for (unsigned i = 0; i != NumPreds; ++i) { 516 // If the value being merged in is not integer or is not defined 517 // in the loop, skip it. 518 Value *InVal = PN->getIncomingValue(i); 519 if (!isa<Instruction>(InVal)) 520 continue; 521 522 // If this pred is for a subloop, not L itself, skip it. 523 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 524 continue; // The Block is in a subloop, skip it. 525 526 // Check that InVal is defined in the loop. 527 Instruction *Inst = cast<Instruction>(InVal); 528 if (!L->contains(Inst)) 529 continue; 530 531 // Okay, this instruction has a user outside of the current loop 532 // and varies predictably *inside* the loop. Evaluate the value it 533 // contains when the loop exits, if possible. 534 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 535 if (!SE->isLoopInvariant(ExitValue, L)) 536 continue; 537 538 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst); 539 540 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 541 << " LoopVal = " << *Inst << "\n"); 542 543 if (!isValidRewrite(Inst, ExitVal)) { 544 DeadInsts.push_back(ExitVal); 545 continue; 546 } 547 Changed = true; 548 ++NumReplaced; 549 550 PN->setIncomingValue(i, ExitVal); 551 552 // If this instruction is dead now, delete it. Don't do it now to avoid 553 // invalidating iterators. 554 if (isInstructionTriviallyDead(Inst, TLI)) 555 DeadInsts.push_back(Inst); 556 557 if (NumPreds == 1) { 558 // Completely replace a single-pred PHI. This is safe, because the 559 // NewVal won't be variant in the loop, so we don't need an LCSSA phi 560 // node anymore. 561 PN->replaceAllUsesWith(ExitVal); 562 PN->eraseFromParent(); 563 } 564 } 565 if (NumPreds != 1) { 566 // Clone the PHI and delete the original one. This lets IVUsers and 567 // any other maps purge the original user from their records. 568 PHINode *NewPN = cast<PHINode>(PN->clone()); 569 NewPN->takeName(PN); 570 NewPN->insertBefore(PN); 571 PN->replaceAllUsesWith(NewPN); 572 PN->eraseFromParent(); 573 } 574 } 575 } 576 577 // The insertion point instruction may have been deleted; clear it out 578 // so that the rewriter doesn't trip over it later. 579 Rewriter.clearInsertPoint(); 580 } 581 582 //===----------------------------------------------------------------------===// 583 // IV Widening - Extend the width of an IV to cover its widest uses. 584 //===----------------------------------------------------------------------===// 585 586 namespace { 587 // Collect information about induction variables that are used by sign/zero 588 // extend operations. This information is recorded by CollectExtend and 589 // provides the input to WidenIV. 590 struct WideIVInfo { 591 PHINode *NarrowIV; 592 Type *WidestNativeType; // Widest integer type created [sz]ext 593 bool IsSigned; // Was an sext user seen before a zext? 594 595 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {} 596 }; 597 598 class WideIVVisitor : public IVVisitor { 599 ScalarEvolution *SE; 600 const DataLayout *TD; 601 602 public: 603 WideIVInfo WI; 604 605 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV, 606 const DataLayout *TData) : 607 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; } 608 609 // Implement the interface used by simplifyUsersOfIV. 610 virtual void visitCast(CastInst *Cast); 611 }; 612 } 613 614 /// visitCast - Update information about the induction variable that is 615 /// extended by this sign or zero extend operation. This is used to determine 616 /// the final width of the IV before actually widening it. 617 void WideIVVisitor::visitCast(CastInst *Cast) { 618 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 619 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 620 return; 621 622 Type *Ty = Cast->getType(); 623 uint64_t Width = SE->getTypeSizeInBits(Ty); 624 if (TD && !TD->isLegalInteger(Width)) 625 return; 626 627 if (!WI.WidestNativeType) { 628 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 629 WI.IsSigned = IsSigned; 630 return; 631 } 632 633 // We extend the IV to satisfy the sign of its first user, arbitrarily. 634 if (WI.IsSigned != IsSigned) 635 return; 636 637 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 638 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 639 } 640 641 namespace { 642 643 /// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the 644 /// WideIV that computes the same value as the Narrow IV def. This avoids 645 /// caching Use* pointers. 646 struct NarrowIVDefUse { 647 Instruction *NarrowDef; 648 Instruction *NarrowUse; 649 Instruction *WideDef; 650 651 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {} 652 653 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD): 654 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {} 655 }; 656 657 /// WidenIV - The goal of this transform is to remove sign and zero extends 658 /// without creating any new induction variables. To do this, it creates a new 659 /// phi of the wider type and redirects all users, either removing extends or 660 /// inserting truncs whenever we stop propagating the type. 661 /// 662 class WidenIV { 663 // Parameters 664 PHINode *OrigPhi; 665 Type *WideType; 666 bool IsSigned; 667 668 // Context 669 LoopInfo *LI; 670 Loop *L; 671 ScalarEvolution *SE; 672 DominatorTree *DT; 673 674 // Result 675 PHINode *WidePhi; 676 Instruction *WideInc; 677 const SCEV *WideIncExpr; 678 SmallVectorImpl<WeakVH> &DeadInsts; 679 680 SmallPtrSet<Instruction*,16> Widened; 681 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 682 683 public: 684 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 685 ScalarEvolution *SEv, DominatorTree *DTree, 686 SmallVectorImpl<WeakVH> &DI) : 687 OrigPhi(WI.NarrowIV), 688 WideType(WI.WidestNativeType), 689 IsSigned(WI.IsSigned), 690 LI(LInfo), 691 L(LI->getLoopFor(OrigPhi->getParent())), 692 SE(SEv), 693 DT(DTree), 694 WidePhi(0), 695 WideInc(0), 696 WideIncExpr(0), 697 DeadInsts(DI) { 698 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 699 } 700 701 PHINode *CreateWideIV(SCEVExpander &Rewriter); 702 703 protected: 704 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 705 Instruction *Use); 706 707 Instruction *CloneIVUser(NarrowIVDefUse DU); 708 709 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse); 710 711 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU); 712 713 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter); 714 715 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 716 }; 717 } // anonymous namespace 718 719 /// isLoopInvariant - Perform a quick domtree based check for loop invariance 720 /// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems 721 /// gratuitous for this purpose. 722 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 723 Instruction *Inst = dyn_cast<Instruction>(V); 724 if (!Inst) 725 return true; 726 727 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 728 } 729 730 Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 731 Instruction *Use) { 732 // Set the debug location and conservative insertion point. 733 IRBuilder<> Builder(Use); 734 // Hoist the insertion point into loop preheaders as far as possible. 735 for (const Loop *L = LI->getLoopFor(Use->getParent()); 736 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 737 L = L->getParentLoop()) 738 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 739 740 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 741 Builder.CreateZExt(NarrowOper, WideType); 742 } 743 744 /// CloneIVUser - Instantiate a wide operation to replace a narrow 745 /// operation. This only needs to handle operations that can evaluation to 746 /// SCEVAddRec. It can safely return 0 for any operation we decide not to clone. 747 Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) { 748 unsigned Opcode = DU.NarrowUse->getOpcode(); 749 switch (Opcode) { 750 default: 751 return 0; 752 case Instruction::Add: 753 case Instruction::Mul: 754 case Instruction::UDiv: 755 case Instruction::Sub: 756 case Instruction::And: 757 case Instruction::Or: 758 case Instruction::Xor: 759 case Instruction::Shl: 760 case Instruction::LShr: 761 case Instruction::AShr: 762 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n"); 763 764 // Replace NarrowDef operands with WideDef. Otherwise, we don't know 765 // anything about the narrow operand yet so must insert a [sz]ext. It is 766 // probably loop invariant and will be folded or hoisted. If it actually 767 // comes from a widened IV, it should be removed during a future call to 768 // WidenIVUse. 769 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef : 770 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse); 771 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef : 772 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse); 773 774 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse); 775 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), 776 LHS, RHS, 777 NarrowBO->getName()); 778 IRBuilder<> Builder(DU.NarrowUse); 779 Builder.Insert(WideBO); 780 if (const OverflowingBinaryOperator *OBO = 781 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) { 782 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap(); 783 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap(); 784 } 785 return WideBO; 786 } 787 } 788 789 /// No-wrap operations can transfer sign extension of their result to their 790 /// operands. Generate the SCEV value for the widened operation without 791 /// actually modifying the IR yet. If the expression after extending the 792 /// operands is an AddRec for this loop, return it. 793 const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) { 794 // Handle the common case of add<nsw/nuw> 795 if (DU.NarrowUse->getOpcode() != Instruction::Add) 796 return 0; 797 798 // One operand (NarrowDef) has already been extended to WideDef. Now determine 799 // if extending the other will lead to a recurrence. 800 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 801 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 802 803 const SCEV *ExtendOperExpr = 0; 804 const OverflowingBinaryOperator *OBO = 805 cast<OverflowingBinaryOperator>(DU.NarrowUse); 806 if (IsSigned && OBO->hasNoSignedWrap()) 807 ExtendOperExpr = SE->getSignExtendExpr( 808 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 809 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 810 ExtendOperExpr = SE->getZeroExtendExpr( 811 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 812 else 813 return 0; 814 815 // When creating this AddExpr, don't apply the current operations NSW or NUW 816 // flags. This instruction may be guarded by control flow that the no-wrap 817 // behavior depends on. Non-control-equivalent instructions can be mapped to 818 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 819 // semantics to those operations. 820 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>( 821 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr)); 822 823 if (!AddRec || AddRec->getLoop() != L) 824 return 0; 825 return AddRec; 826 } 827 828 /// GetWideRecurrence - Is this instruction potentially interesting from 829 /// IVUsers' perspective after widening it's type? In other words, can the 830 /// extend be safely hoisted out of the loop with SCEV reducing the value to a 831 /// recurrence on the same loop. If so, return the sign or zero extended 832 /// recurrence. Otherwise return NULL. 833 const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) { 834 if (!SE->isSCEVable(NarrowUse->getType())) 835 return 0; 836 837 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 838 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 839 >= SE->getTypeSizeInBits(WideType)) { 840 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 841 // index. So don't follow this use. 842 return 0; 843 } 844 845 const SCEV *WideExpr = IsSigned ? 846 SE->getSignExtendExpr(NarrowExpr, WideType) : 847 SE->getZeroExtendExpr(NarrowExpr, WideType); 848 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 849 if (!AddRec || AddRec->getLoop() != L) 850 return 0; 851 return AddRec; 852 } 853 854 /// WidenIVUse - Determine whether an individual user of the narrow IV can be 855 /// widened. If so, return the wide clone of the user. 856 Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) { 857 858 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 859 if (isa<PHINode>(DU.NarrowUse) && 860 LI->getLoopFor(DU.NarrowUse->getParent()) != L) 861 return 0; 862 863 // Our raison d'etre! Eliminate sign and zero extension. 864 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 865 Value *NewDef = DU.WideDef; 866 if (DU.NarrowUse->getType() != WideType) { 867 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 868 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 869 if (CastWidth < IVWidth) { 870 // The cast isn't as wide as the IV, so insert a Trunc. 871 IRBuilder<> Builder(DU.NarrowUse); 872 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 873 } 874 else { 875 // A wider extend was hidden behind a narrower one. This may induce 876 // another round of IV widening in which the intermediate IV becomes 877 // dead. It should be very rare. 878 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 879 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 880 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 881 NewDef = DU.NarrowUse; 882 } 883 } 884 if (NewDef != DU.NarrowUse) { 885 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 886 << " replaced by " << *DU.WideDef << "\n"); 887 ++NumElimExt; 888 DU.NarrowUse->replaceAllUsesWith(NewDef); 889 DeadInsts.push_back(DU.NarrowUse); 890 } 891 // Now that the extend is gone, we want to expose it's uses for potential 892 // further simplification. We don't need to directly inform SimplifyIVUsers 893 // of the new users, because their parent IV will be processed later as a 894 // new loop phi. If we preserved IVUsers analysis, we would also want to 895 // push the uses of WideDef here. 896 897 // No further widening is needed. The deceased [sz]ext had done it for us. 898 return 0; 899 } 900 901 // Does this user itself evaluate to a recurrence after widening? 902 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse); 903 if (!WideAddRec) { 904 WideAddRec = GetExtendedOperandRecurrence(DU); 905 } 906 if (!WideAddRec) { 907 // This user does not evaluate to a recurence after widening, so don't 908 // follow it. Instead insert a Trunc to kill off the original use, 909 // eventually isolating the original narrow IV so it can be removed. 910 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT)); 911 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 912 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 913 return 0; 914 } 915 // Assume block terminators cannot evaluate to a recurrence. We can't to 916 // insert a Trunc after a terminator if there happens to be a critical edge. 917 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 918 "SCEV is not expected to evaluate a block terminator"); 919 920 // Reuse the IV increment that SCEVExpander created as long as it dominates 921 // NarrowUse. 922 Instruction *WideUse = 0; 923 if (WideAddRec == WideIncExpr 924 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse)) 925 WideUse = WideInc; 926 else { 927 WideUse = CloneIVUser(DU); 928 if (!WideUse) 929 return 0; 930 } 931 // Evaluation of WideAddRec ensured that the narrow expression could be 932 // extended outside the loop without overflow. This suggests that the wide use 933 // evaluates to the same expression as the extended narrow use, but doesn't 934 // absolutely guarantee it. Hence the following failsafe check. In rare cases 935 // where it fails, we simply throw away the newly created wide use. 936 if (WideAddRec != SE->getSCEV(WideUse)) { 937 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 938 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 939 DeadInsts.push_back(WideUse); 940 return 0; 941 } 942 943 // Returning WideUse pushes it on the worklist. 944 return WideUse; 945 } 946 947 /// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers. 948 /// 949 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 950 for (Value::use_iterator UI = NarrowDef->use_begin(), 951 UE = NarrowDef->use_end(); UI != UE; ++UI) { 952 Instruction *NarrowUse = cast<Instruction>(*UI); 953 954 // Handle data flow merges and bizarre phi cycles. 955 if (!Widened.insert(NarrowUse)) 956 continue; 957 958 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef)); 959 } 960 } 961 962 /// CreateWideIV - Process a single induction variable. First use the 963 /// SCEVExpander to create a wide induction variable that evaluates to the same 964 /// recurrence as the original narrow IV. Then use a worklist to forward 965 /// traverse the narrow IV's def-use chain. After WidenIVUse has processed all 966 /// interesting IV users, the narrow IV will be isolated for removal by 967 /// DeleteDeadPHIs. 968 /// 969 /// It would be simpler to delete uses as they are processed, but we must avoid 970 /// invalidating SCEV expressions. 971 /// 972 PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) { 973 // Is this phi an induction variable? 974 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 975 if (!AddRec) 976 return NULL; 977 978 // Widen the induction variable expression. 979 const SCEV *WideIVExpr = IsSigned ? 980 SE->getSignExtendExpr(AddRec, WideType) : 981 SE->getZeroExtendExpr(AddRec, WideType); 982 983 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 984 "Expect the new IV expression to preserve its type"); 985 986 // Can the IV be extended outside the loop without overflow? 987 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 988 if (!AddRec || AddRec->getLoop() != L) 989 return NULL; 990 991 // An AddRec must have loop-invariant operands. Since this AddRec is 992 // materialized by a loop header phi, the expression cannot have any post-loop 993 // operands, so they must dominate the loop header. 994 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 995 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) 996 && "Loop header phi recurrence inputs do not dominate the loop"); 997 998 // The rewriter provides a value for the desired IV expression. This may 999 // either find an existing phi or materialize a new one. Either way, we 1000 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1001 // of the phi-SCC dominates the loop entry. 1002 Instruction *InsertPt = L->getHeader()->begin(); 1003 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1004 1005 // Remembering the WideIV increment generated by SCEVExpander allows 1006 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't 1007 // employ a general reuse mechanism because the call above is the only call to 1008 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1009 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1010 WideInc = 1011 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1012 WideIncExpr = SE->getSCEV(WideInc); 1013 } 1014 1015 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1016 ++NumWidened; 1017 1018 // Traverse the def-use chain using a worklist starting at the original IV. 1019 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1020 1021 Widened.insert(OrigPhi); 1022 pushNarrowIVUsers(OrigPhi, WidePhi); 1023 1024 while (!NarrowIVUsers.empty()) { 1025 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1026 1027 // Process a def-use edge. This may replace the use, so don't hold a 1028 // use_iterator across it. 1029 Instruction *WideUse = WidenIVUse(DU, Rewriter); 1030 1031 // Follow all def-use edges from the previous narrow use. 1032 if (WideUse) 1033 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1034 1035 // WidenIVUse may have removed the def-use edge. 1036 if (DU.NarrowDef->use_empty()) 1037 DeadInsts.push_back(DU.NarrowDef); 1038 } 1039 return WidePhi; 1040 } 1041 1042 //===----------------------------------------------------------------------===// 1043 // Simplification of IV users based on SCEV evaluation. 1044 //===----------------------------------------------------------------------===// 1045 1046 1047 /// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV 1048 /// users. Each successive simplification may push more users which may 1049 /// themselves be candidates for simplification. 1050 /// 1051 /// Sign/Zero extend elimination is interleaved with IV simplification. 1052 /// 1053 void IndVarSimplify::SimplifyAndExtend(Loop *L, 1054 SCEVExpander &Rewriter, 1055 LPPassManager &LPM) { 1056 SmallVector<WideIVInfo, 8> WideIVs; 1057 1058 SmallVector<PHINode*, 8> LoopPhis; 1059 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1060 LoopPhis.push_back(cast<PHINode>(I)); 1061 } 1062 // Each round of simplification iterates through the SimplifyIVUsers worklist 1063 // for all current phis, then determines whether any IVs can be 1064 // widened. Widening adds new phis to LoopPhis, inducing another round of 1065 // simplification on the wide IVs. 1066 while (!LoopPhis.empty()) { 1067 // Evaluate as many IV expressions as possible before widening any IVs. This 1068 // forces SCEV to set no-wrap flags before evaluating sign/zero 1069 // extension. The first time SCEV attempts to normalize sign/zero extension, 1070 // the result becomes final. So for the most predictable results, we delay 1071 // evaluation of sign/zero extend evaluation until needed, and avoid running 1072 // other SCEV based analysis prior to SimplifyAndExtend. 1073 do { 1074 PHINode *CurrIV = LoopPhis.pop_back_val(); 1075 1076 // Information about sign/zero extensions of CurrIV. 1077 WideIVVisitor WIV(CurrIV, SE, TD); 1078 1079 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV); 1080 1081 if (WIV.WI.WidestNativeType) { 1082 WideIVs.push_back(WIV.WI); 1083 } 1084 } while(!LoopPhis.empty()); 1085 1086 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1087 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1088 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) { 1089 Changed = true; 1090 LoopPhis.push_back(WidePhi); 1091 } 1092 } 1093 } 1094 } 1095 1096 //===----------------------------------------------------------------------===// 1097 // LinearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1098 //===----------------------------------------------------------------------===// 1099 1100 /// Check for expressions that ScalarEvolution generates to compute 1101 /// BackedgeTakenInfo. If these expressions have not been reduced, then 1102 /// expanding them may incur additional cost (albeit in the loop preheader). 1103 static bool isHighCostExpansion(const SCEV *S, BranchInst *BI, 1104 SmallPtrSet<const SCEV*, 8> &Processed, 1105 ScalarEvolution *SE) { 1106 if (!Processed.insert(S)) 1107 return false; 1108 1109 // If the backedge-taken count is a UDiv, it's very likely a UDiv that 1110 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a 1111 // precise expression, rather than a UDiv from the user's code. If we can't 1112 // find a UDiv in the code with some simple searching, assume the former and 1113 // forego rewriting the loop. 1114 if (isa<SCEVUDivExpr>(S)) { 1115 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition()); 1116 if (!OrigCond) return true; 1117 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1)); 1118 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1)); 1119 if (R != S) { 1120 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0)); 1121 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1)); 1122 if (L != S) 1123 return true; 1124 } 1125 } 1126 1127 // Recurse past add expressions, which commonly occur in the 1128 // BackedgeTakenCount. They may already exist in program code, and if not, 1129 // they are not too expensive rematerialize. 1130 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 1131 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 1132 I != E; ++I) { 1133 if (isHighCostExpansion(*I, BI, Processed, SE)) 1134 return true; 1135 } 1136 return false; 1137 } 1138 1139 // HowManyLessThans uses a Max expression whenever the loop is not guarded by 1140 // the exit condition. 1141 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 1142 return true; 1143 1144 // If we haven't recognized an expensive SCEV pattern, assume it's an 1145 // expression produced by program code. 1146 return false; 1147 } 1148 1149 /// canExpandBackedgeTakenCount - Return true if this loop's backedge taken 1150 /// count expression can be safely and cheaply expanded into an instruction 1151 /// sequence that can be used by LinearFunctionTestReplace. 1152 /// 1153 /// TODO: This fails for pointer-type loop counters with greater than one byte 1154 /// strides, consequently preventing LFTR from running. For the purpose of LFTR 1155 /// we could skip this check in the case that the LFTR loop counter (chosen by 1156 /// FindLoopCounter) is also pointer type. Instead, we could directly convert 1157 /// the loop test to an inequality test by checking the target data's alignment 1158 /// of element types (given that the initial pointer value originates from or is 1159 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1160 /// However, we don't yet have a strong motivation for converting loop tests 1161 /// into inequality tests. 1162 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) { 1163 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1164 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1165 BackedgeTakenCount->isZero()) 1166 return false; 1167 1168 if (!L->getExitingBlock()) 1169 return false; 1170 1171 // Can't rewrite non-branch yet. 1172 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1173 if (!BI) 1174 return false; 1175 1176 SmallPtrSet<const SCEV*, 8> Processed; 1177 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE)) 1178 return false; 1179 1180 return true; 1181 } 1182 1183 /// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop 1184 /// invariant value to the phi. 1185 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1186 Instruction *IncI = dyn_cast<Instruction>(IncV); 1187 if (!IncI) 1188 return 0; 1189 1190 switch (IncI->getOpcode()) { 1191 case Instruction::Add: 1192 case Instruction::Sub: 1193 break; 1194 case Instruction::GetElementPtr: 1195 // An IV counter must preserve its type. 1196 if (IncI->getNumOperands() == 2) 1197 break; 1198 default: 1199 return 0; 1200 } 1201 1202 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1203 if (Phi && Phi->getParent() == L->getHeader()) { 1204 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1205 return Phi; 1206 return 0; 1207 } 1208 if (IncI->getOpcode() == Instruction::GetElementPtr) 1209 return 0; 1210 1211 // Allow add/sub to be commuted. 1212 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1213 if (Phi && Phi->getParent() == L->getHeader()) { 1214 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1215 return Phi; 1216 } 1217 return 0; 1218 } 1219 1220 /// Return the compare guarding the loop latch, or NULL for unrecognized tests. 1221 static ICmpInst *getLoopTest(Loop *L) { 1222 assert(L->getExitingBlock() && "expected loop exit"); 1223 1224 BasicBlock *LatchBlock = L->getLoopLatch(); 1225 // Don't bother with LFTR if the loop is not properly simplified. 1226 if (!LatchBlock) 1227 return 0; 1228 1229 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1230 assert(BI && "expected exit branch"); 1231 1232 return dyn_cast<ICmpInst>(BI->getCondition()); 1233 } 1234 1235 /// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show 1236 /// that the current exit test is already sufficiently canonical. 1237 static bool needsLFTR(Loop *L, DominatorTree *DT) { 1238 // Do LFTR to simplify the exit condition to an ICMP. 1239 ICmpInst *Cond = getLoopTest(L); 1240 if (!Cond) 1241 return true; 1242 1243 // Do LFTR to simplify the exit ICMP to EQ/NE 1244 ICmpInst::Predicate Pred = Cond->getPredicate(); 1245 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1246 return true; 1247 1248 // Look for a loop invariant RHS 1249 Value *LHS = Cond->getOperand(0); 1250 Value *RHS = Cond->getOperand(1); 1251 if (!isLoopInvariant(RHS, L, DT)) { 1252 if (!isLoopInvariant(LHS, L, DT)) 1253 return true; 1254 std::swap(LHS, RHS); 1255 } 1256 // Look for a simple IV counter LHS 1257 PHINode *Phi = dyn_cast<PHINode>(LHS); 1258 if (!Phi) 1259 Phi = getLoopPhiForCounter(LHS, L, DT); 1260 1261 if (!Phi) 1262 return true; 1263 1264 // Do LFTR if PHI node is defined in the loop, but is *not* a counter. 1265 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); 1266 if (Idx < 0) 1267 return true; 1268 1269 // Do LFTR if the exit condition's IV is *not* a simple counter. 1270 Value *IncV = Phi->getIncomingValue(Idx); 1271 return Phi != getLoopPhiForCounter(IncV, L, DT); 1272 } 1273 1274 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils 1275 /// down to checking that all operands are constant and listing instructions 1276 /// that may hide undef. 1277 static bool hasConcreteDefImpl(Value *V, SmallPtrSet<Value*, 8> &Visited, 1278 unsigned Depth) { 1279 if (isa<Constant>(V)) 1280 return !isa<UndefValue>(V); 1281 1282 if (Depth >= 6) 1283 return false; 1284 1285 // Conservatively handle non-constant non-instructions. For example, Arguments 1286 // may be undef. 1287 Instruction *I = dyn_cast<Instruction>(V); 1288 if (!I) 1289 return false; 1290 1291 // Load and return values may be undef. 1292 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I)) 1293 return false; 1294 1295 // Optimistically handle other instructions. 1296 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) { 1297 if (!Visited.insert(*OI)) 1298 continue; 1299 if (!hasConcreteDefImpl(*OI, Visited, Depth+1)) 1300 return false; 1301 } 1302 return true; 1303 } 1304 1305 /// Return true if the given value is concrete. We must prove that undef can 1306 /// never reach it. 1307 /// 1308 /// TODO: If we decide that this is a good approach to checking for undef, we 1309 /// may factor it into a common location. 1310 static bool hasConcreteDef(Value *V) { 1311 SmallPtrSet<Value*, 8> Visited; 1312 Visited.insert(V); 1313 return hasConcreteDefImpl(V, Visited, 0); 1314 } 1315 1316 /// AlmostDeadIV - Return true if this IV has any uses other than the (soon to 1317 /// be rewritten) loop exit test. 1318 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1319 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1320 Value *IncV = Phi->getIncomingValue(LatchIdx); 1321 1322 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end(); 1323 UI != UE; ++UI) { 1324 if (*UI != Cond && *UI != IncV) return false; 1325 } 1326 1327 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end(); 1328 UI != UE; ++UI) { 1329 if (*UI != Cond && *UI != Phi) return false; 1330 } 1331 return true; 1332 } 1333 1334 /// FindLoopCounter - Find an affine IV in canonical form. 1335 /// 1336 /// BECount may be an i8* pointer type. The pointer difference is already 1337 /// valid count without scaling the address stride, so it remains a pointer 1338 /// expression as far as SCEV is concerned. 1339 /// 1340 /// Currently only valid for LFTR. See the comments on hasConcreteDef below. 1341 /// 1342 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1343 /// 1344 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1345 /// This is difficult in general for SCEV because of potential overflow. But we 1346 /// could at least handle constant BECounts. 1347 static PHINode * 1348 FindLoopCounter(Loop *L, const SCEV *BECount, 1349 ScalarEvolution *SE, DominatorTree *DT, const DataLayout *TD) { 1350 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1351 1352 Value *Cond = 1353 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1354 1355 // Loop over all of the PHI nodes, looking for a simple counter. 1356 PHINode *BestPhi = 0; 1357 const SCEV *BestInit = 0; 1358 BasicBlock *LatchBlock = L->getLoopLatch(); 1359 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1360 1361 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1362 PHINode *Phi = cast<PHINode>(I); 1363 if (!SE->isSCEVable(Phi->getType())) 1364 continue; 1365 1366 // Avoid comparing an integer IV against a pointer Limit. 1367 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1368 continue; 1369 1370 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1371 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1372 continue; 1373 1374 // AR may be a pointer type, while BECount is an integer type. 1375 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1376 // AR may not be a narrower type, or we may never exit. 1377 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1378 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth))) 1379 continue; 1380 1381 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1382 if (!Step || !Step->isOne()) 1383 continue; 1384 1385 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1386 Value *IncV = Phi->getIncomingValue(LatchIdx); 1387 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1388 continue; 1389 1390 // Avoid reusing a potentially undef value to compute other values that may 1391 // have originally had a concrete definition. 1392 if (!hasConcreteDef(Phi)) { 1393 // We explicitly allow unknown phis as long as they are already used by 1394 // the loop test. In this case we assume that performing LFTR could not 1395 // increase the number of undef users. 1396 if (ICmpInst *Cond = getLoopTest(L)) { 1397 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) 1398 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) { 1399 continue; 1400 } 1401 } 1402 } 1403 const SCEV *Init = AR->getStart(); 1404 1405 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1406 // Don't force a live loop counter if another IV can be used. 1407 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1408 continue; 1409 1410 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1411 // also prefers integer to pointer IVs. 1412 if (BestInit->isZero() != Init->isZero()) { 1413 if (BestInit->isZero()) 1414 continue; 1415 } 1416 // If two IVs both count from zero or both count from nonzero then the 1417 // narrower is likely a dead phi that has been widened. Use the wider phi 1418 // to allow the other to be eliminated. 1419 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1420 continue; 1421 } 1422 BestPhi = Phi; 1423 BestInit = Init; 1424 } 1425 return BestPhi; 1426 } 1427 1428 /// genLoopLimit - Help LinearFunctionTestReplace by generating a value that 1429 /// holds the RHS of the new loop test. 1430 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1431 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1432 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1433 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1434 const SCEV *IVInit = AR->getStart(); 1435 1436 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1437 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1438 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1439 // the existing GEPs whenever possible. 1440 if (IndVar->getType()->isPointerTy() 1441 && !IVCount->getType()->isPointerTy()) { 1442 1443 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1444 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy); 1445 1446 // Expand the code for the iteration count. 1447 assert(SE->isLoopInvariant(IVOffset, L) && 1448 "Computed iteration count is not loop invariant!"); 1449 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1450 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1451 1452 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1453 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1454 // We could handle pointer IVs other than i8*, but we need to compensate for 1455 // gep index scaling. See canExpandBackedgeTakenCount comments. 1456 assert(SE->getSizeOfExpr( 1457 cast<PointerType>(GEPBase->getType())->getElementType())->isOne() 1458 && "unit stride pointer IV must be i8*"); 1459 1460 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1461 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit"); 1462 } 1463 else { 1464 // In any other case, convert both IVInit and IVCount to integers before 1465 // comparing. This may result in SCEV expension of pointers, but in practice 1466 // SCEV will fold the pointer arithmetic away as such: 1467 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1468 // 1469 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1470 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a 1471 // pointer may occur when enable-iv-rewrite generates a canonical IV on top 1472 // of case #2. 1473 1474 const SCEV *IVLimit = 0; 1475 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1476 // For non-zero Start, compute IVCount here. 1477 if (AR->getStart()->isZero()) 1478 IVLimit = IVCount; 1479 else { 1480 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1481 const SCEV *IVInit = AR->getStart(); 1482 1483 // For integer IVs, truncate the IV before computing IVInit + BECount. 1484 if (SE->getTypeSizeInBits(IVInit->getType()) 1485 > SE->getTypeSizeInBits(IVCount->getType())) 1486 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1487 1488 IVLimit = SE->getAddExpr(IVInit, IVCount); 1489 } 1490 // Expand the code for the iteration count. 1491 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1492 IRBuilder<> Builder(BI); 1493 assert(SE->isLoopInvariant(IVLimit, L) && 1494 "Computed iteration count is not loop invariant!"); 1495 // Ensure that we generate the same type as IndVar, or a smaller integer 1496 // type. In the presence of null pointer values, we have an integer type 1497 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1498 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1499 IndVar->getType() : IVCount->getType(); 1500 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1501 } 1502 } 1503 1504 /// LinearFunctionTestReplace - This method rewrites the exit condition of the 1505 /// loop to be a canonical != comparison against the incremented loop induction 1506 /// variable. This pass is able to rewrite the exit tests of any loop where the 1507 /// SCEV analysis can determine a loop-invariant trip count of the loop, which 1508 /// is actually a much broader range than just linear tests. 1509 Value *IndVarSimplify:: 1510 LinearFunctionTestReplace(Loop *L, 1511 const SCEV *BackedgeTakenCount, 1512 PHINode *IndVar, 1513 SCEVExpander &Rewriter) { 1514 assert(canExpandBackedgeTakenCount(L, SE) && "precondition"); 1515 1516 // LFTR can ignore IV overflow and truncate to the width of 1517 // BECount. This avoids materializing the add(zext(add)) expression. 1518 Type *CntTy = BackedgeTakenCount->getType(); 1519 1520 const SCEV *IVCount = BackedgeTakenCount; 1521 1522 // If the exiting block is the same as the backedge block, we prefer to 1523 // compare against the post-incremented value, otherwise we must compare 1524 // against the preincremented value. 1525 Value *CmpIndVar; 1526 if (L->getExitingBlock() == L->getLoopLatch()) { 1527 // Add one to the "backedge-taken" count to get the trip count. 1528 // If this addition may overflow, we have to be more pessimistic and 1529 // cast the induction variable before doing the add. 1530 const SCEV *N = 1531 SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1)); 1532 if (CntTy == IVCount->getType()) 1533 IVCount = N; 1534 else { 1535 const SCEV *Zero = SE->getConstant(IVCount->getType(), 0); 1536 if ((isa<SCEVConstant>(N) && !N->isZero()) || 1537 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) { 1538 // No overflow. Cast the sum. 1539 IVCount = SE->getTruncateOrZeroExtend(N, CntTy); 1540 } else { 1541 // Potential overflow. Cast before doing the add. 1542 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy); 1543 IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1)); 1544 } 1545 } 1546 // The BackedgeTaken expression contains the number of times that the 1547 // backedge branches to the loop header. This is one less than the 1548 // number of times the loop executes, so use the incremented indvar. 1549 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1550 } else { 1551 // We must use the preincremented value... 1552 IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy); 1553 CmpIndVar = IndVar; 1554 } 1555 1556 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1557 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy() 1558 && "genLoopLimit missed a cast"); 1559 1560 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1561 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1562 ICmpInst::Predicate P; 1563 if (L->contains(BI->getSuccessor(0))) 1564 P = ICmpInst::ICMP_NE; 1565 else 1566 P = ICmpInst::ICMP_EQ; 1567 1568 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1569 << " LHS:" << *CmpIndVar << '\n' 1570 << " op:\t" 1571 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1572 << " RHS:\t" << *ExitCnt << "\n" 1573 << " IVCount:\t" << *IVCount << "\n"); 1574 1575 IRBuilder<> Builder(BI); 1576 if (SE->getTypeSizeInBits(CmpIndVar->getType()) 1577 > SE->getTypeSizeInBits(ExitCnt->getType())) { 1578 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1579 "lftr.wideiv"); 1580 } 1581 1582 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 1583 Value *OrigCond = BI->getCondition(); 1584 // It's tempting to use replaceAllUsesWith here to fully replace the old 1585 // comparison, but that's not immediately safe, since users of the old 1586 // comparison may not be dominated by the new comparison. Instead, just 1587 // update the branch to use the new comparison; in the common case this 1588 // will make old comparison dead. 1589 BI->setCondition(Cond); 1590 DeadInsts.push_back(OrigCond); 1591 1592 ++NumLFTR; 1593 Changed = true; 1594 return Cond; 1595 } 1596 1597 //===----------------------------------------------------------------------===// 1598 // SinkUnusedInvariants. A late subpass to cleanup loop preheaders. 1599 //===----------------------------------------------------------------------===// 1600 1601 /// If there's a single exit block, sink any loop-invariant values that 1602 /// were defined in the preheader but not used inside the loop into the 1603 /// exit block to reduce register pressure in the loop. 1604 void IndVarSimplify::SinkUnusedInvariants(Loop *L) { 1605 BasicBlock *ExitBlock = L->getExitBlock(); 1606 if (!ExitBlock) return; 1607 1608 BasicBlock *Preheader = L->getLoopPreheader(); 1609 if (!Preheader) return; 1610 1611 Instruction *InsertPt = ExitBlock->getFirstInsertionPt(); 1612 BasicBlock::iterator I = Preheader->getTerminator(); 1613 while (I != Preheader->begin()) { 1614 --I; 1615 // New instructions were inserted at the end of the preheader. 1616 if (isa<PHINode>(I)) 1617 break; 1618 1619 // Don't move instructions which might have side effects, since the side 1620 // effects need to complete before instructions inside the loop. Also don't 1621 // move instructions which might read memory, since the loop may modify 1622 // memory. Note that it's okay if the instruction might have undefined 1623 // behavior: LoopSimplify guarantees that the preheader dominates the exit 1624 // block. 1625 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 1626 continue; 1627 1628 // Skip debug info intrinsics. 1629 if (isa<DbgInfoIntrinsic>(I)) 1630 continue; 1631 1632 // Skip landingpad instructions. 1633 if (isa<LandingPadInst>(I)) 1634 continue; 1635 1636 // Don't sink alloca: we never want to sink static alloca's out of the 1637 // entry block, and correctly sinking dynamic alloca's requires 1638 // checks for stacksave/stackrestore intrinsics. 1639 // FIXME: Refactor this check somehow? 1640 if (isa<AllocaInst>(I)) 1641 continue; 1642 1643 // Determine if there is a use in or before the loop (direct or 1644 // otherwise). 1645 bool UsedInLoop = false; 1646 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); 1647 UI != UE; ++UI) { 1648 User *U = *UI; 1649 BasicBlock *UseBB = cast<Instruction>(U)->getParent(); 1650 if (PHINode *P = dyn_cast<PHINode>(U)) { 1651 unsigned i = 1652 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()); 1653 UseBB = P->getIncomingBlock(i); 1654 } 1655 if (UseBB == Preheader || L->contains(UseBB)) { 1656 UsedInLoop = true; 1657 break; 1658 } 1659 } 1660 1661 // If there is, the def must remain in the preheader. 1662 if (UsedInLoop) 1663 continue; 1664 1665 // Otherwise, sink it to the exit block. 1666 Instruction *ToMove = I; 1667 bool Done = false; 1668 1669 if (I != Preheader->begin()) { 1670 // Skip debug info intrinsics. 1671 do { 1672 --I; 1673 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 1674 1675 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 1676 Done = true; 1677 } else { 1678 Done = true; 1679 } 1680 1681 ToMove->moveBefore(InsertPt); 1682 if (Done) break; 1683 InsertPt = ToMove; 1684 } 1685 } 1686 1687 //===----------------------------------------------------------------------===// 1688 // IndVarSimplify driver. Manage several subpasses of IV simplification. 1689 //===----------------------------------------------------------------------===// 1690 1691 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { 1692 // If LoopSimplify form is not available, stay out of trouble. Some notes: 1693 // - LSR currently only supports LoopSimplify-form loops. Indvars' 1694 // canonicalization can be a pessimization without LSR to "clean up" 1695 // afterwards. 1696 // - We depend on having a preheader; in particular, 1697 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 1698 // and we're in trouble if we can't find the induction variable even when 1699 // we've manually inserted one. 1700 if (!L->isLoopSimplifyForm()) 1701 return false; 1702 1703 LI = &getAnalysis<LoopInfo>(); 1704 SE = &getAnalysis<ScalarEvolution>(); 1705 DT = &getAnalysis<DominatorTree>(); 1706 TD = getAnalysisIfAvailable<DataLayout>(); 1707 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 1708 1709 DeadInsts.clear(); 1710 Changed = false; 1711 1712 // If there are any floating-point recurrences, attempt to 1713 // transform them to use integer recurrences. 1714 RewriteNonIntegerIVs(L); 1715 1716 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1717 1718 // Create a rewriter object which we'll use to transform the code with. 1719 SCEVExpander Rewriter(*SE, "indvars"); 1720 #ifndef NDEBUG 1721 Rewriter.setDebugType(DEBUG_TYPE); 1722 #endif 1723 1724 // Eliminate redundant IV users. 1725 // 1726 // Simplification works best when run before other consumers of SCEV. We 1727 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 1728 // other expressions involving loop IVs have been evaluated. This helps SCEV 1729 // set no-wrap flags before normalizing sign/zero extension. 1730 Rewriter.disableCanonicalMode(); 1731 SimplifyAndExtend(L, Rewriter, LPM); 1732 1733 // Check to see if this loop has a computable loop-invariant execution count. 1734 // If so, this means that we can compute the final value of any expressions 1735 // that are recurrent in the loop, and substitute the exit values from the 1736 // loop into any instructions outside of the loop that use the final values of 1737 // the current expressions. 1738 // 1739 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1740 RewriteLoopExitValues(L, Rewriter); 1741 1742 // Eliminate redundant IV cycles. 1743 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 1744 1745 // If we have a trip count expression, rewrite the loop's exit condition 1746 // using it. We can currently only handle loops with a single exit. 1747 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) { 1748 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD); 1749 if (IndVar) { 1750 // Check preconditions for proper SCEVExpander operation. SCEV does not 1751 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 1752 // pass that uses the SCEVExpander must do it. This does not work well for 1753 // loop passes because SCEVExpander makes assumptions about all loops, while 1754 // LoopPassManager only forces the current loop to be simplified. 1755 // 1756 // FIXME: SCEV expansion has no way to bail out, so the caller must 1757 // explicitly check any assumptions made by SCEV. Brittle. 1758 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 1759 if (!AR || AR->getLoop()->getLoopPreheader()) 1760 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, 1761 Rewriter); 1762 } 1763 } 1764 // Clear the rewriter cache, because values that are in the rewriter's cache 1765 // can be deleted in the loop below, causing the AssertingVH in the cache to 1766 // trigger. 1767 Rewriter.clear(); 1768 1769 // Now that we're done iterating through lists, clean up any instructions 1770 // which are now dead. 1771 while (!DeadInsts.empty()) 1772 if (Instruction *Inst = 1773 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) 1774 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 1775 1776 // The Rewriter may not be used from this point on. 1777 1778 // Loop-invariant instructions in the preheader that aren't used in the 1779 // loop may be sunk below the loop to reduce register pressure. 1780 SinkUnusedInvariants(L); 1781 1782 // Clean up dead instructions. 1783 Changed |= DeleteDeadPHIs(L->getHeader(), TLI); 1784 // Check a post-condition. 1785 assert(L->isLCSSAForm(*DT) && 1786 "Indvars did not leave the loop in lcssa form!"); 1787 1788 // Verify that LFTR, and any other change have not interfered with SCEV's 1789 // ability to compute trip count. 1790 #ifndef NDEBUG 1791 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 1792 SE->forgetLoop(L); 1793 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 1794 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 1795 SE->getTypeSizeInBits(NewBECount->getType())) 1796 NewBECount = SE->getTruncateOrNoop(NewBECount, 1797 BackedgeTakenCount->getType()); 1798 else 1799 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 1800 NewBECount->getType()); 1801 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 1802 } 1803 #endif 1804 1805 return Changed; 1806 } 1807